netprojects/JEV-35B-mlx-8bit

JEV-35B: MLX 8-bit

A 8-bit MLX build of autotrust/JEV-35B for Apple Silicon, with vision. 36 GB. Released 19 October 2026.

JEV-35B is AutoTrust AI's System 1 decision model, a fine-tune of Qwen3.5-35B-A3B (35 B parameters, about 3 B active per token). Each call is one typed question about a state (text, JSON or images), answered in one forward pass with a calibrated probability for every allowed option: choice (2 to 256 options you supply), noul (yes / no) or score (0 to 5). It classifies, routes, gates and ranks; it does not chat or generate text.

This build: Every layer at 8 bits (affine, group size 64), except the vision tower and the router gates, which stay in bf16. Recommended for Macs with 48 GB or more. Effectively the original model: it picks the same answer on 228 of 231 decisions and its probabilities drift by a KL of 0.002 on average.

Published by NetProjects (Internet Projects Ltd). Not affiliated with AutoTrust AI.

Use

pip install "git+https://github.com/internet-projects/jev35b-mlx"

import jev_mlx

jev = jev_mlx.load("netprojects/JEV-35B-mlx-8bit")
r = jev.decide("choice",
               state="Customer message: my card was charged twice for the same order.",
               question="Which team should handle this ticket?",
               options=["billing", "shipping", "technical support", "account security"])
print(r["choice"], r["probabilities"])

r = jev.decide("noul", state=["Photo from the warehouse:", {"image": "shelf.jpg"}],
               question="Is any box on the shelf damaged?")
print(r["probabilities"][1])      # P(true)

Or as a server with the same /v1/decide fields as autotrust's own:

python -m jev_mlx serve netprojects/JEV-35B-mlx-8bit --port 8000
curl localhost:8000/v1/decide -H 'Content-Type: application/json' -d '{"kind": "noul",
  "state": "Order 1182 arrived with a cracked screen.", "question": "Does the customer report damage?"}'

jev35b-mlx reproduces the read-out of autotrust's serve_decide.py: the answer-slot logits at the last position, plus the decision-head bias, divided by the per-kind temperature in calibration.json. Both files ship in this repository.

Verified against the original weights

231 decisions built from the JevBench public set, each compared with the original bf16 weights (converted to MLX without quantization) on a MacBook Pro M5 Max:

build	size	memory in use	same top answer	probability drift, KL mean / p95	warm median (p95) per decision
8-bit	36 GB	37.7 GB	228 / 231	0.002 / 0.009	164 ms (1,037 ms)
4-bit mixed	21 GB	21.6 GB	225 / 231	0.004 / 0.021	109 ms (246 ms)
3-bit mixed	17 GB	17.6 GB	224 / 231	0.012 / 0.056	99 ms (234 ms)

• Images: every build answers our vision check (two generated images, a choice and a yes/no question each) 4 / 4.
• Speed: one request at a time, prompts of about 94 tokens, with other models running on the same Mac. The first requests after loading take a few hundred ms longer while the experts are read from disk.
• In size, this build sits next to the Q8_0 GGUF; the vision tower is included in the 36 GB.
• The scripts to reproduce these numbers are in jev35b-mlx (scripts/check.py, scripts/vision_check.py, scripts/speed.py).

For the model's own benchmarks (Decision Index, vision, ERQA, computer use), see the original model card.

More from NetProjects

• jev35b-mlx on GitHub and the NetProjects Hugging Face organization
• H2O-Lightning-4B on MLX, another decision model
• NVIDIA Nemotron-Labs Dory 1.0 4B on MLX

Credits and license

Model: autotrust/JEV-35B by AutoTrust AI (Apache-2.0, the same license as the original), a fine-tune of Qwen/Qwen3.5-35B-A3B (Apache-2.0). Converted with mlx-vlm. Code in jev35b-mlx: Apache-2.0. Port by Paris Michalas, Partner, NetProjects (Internet Projects Ltd). Not affiliated with AutoTrust AI.